Security Testing. Ulf Kargén Department of Computer and Information Science (IDA) Division for Database and Information Techniques (ADIT)

Transcription

1 Security Testing TDDC90 Software Security Ulf Kargén Department of Computer and Information Science (IDA) Division for Database and Information Techniques (ADIT)

2 Security testing vs regular testing Regular testing aims to ensure that the program meets customer requirements in terms of features and functionality. Tests normal use cases Test with regards to common expected usage patterns. Security testing aims to ensure that program fulfills security requirements. Often non-functional. More interested in misuse cases Attackers taking advantage of weird corner cases. 2

3 Functional vs non-functional security requirements Functional requirements What shall the software do? Non-functional requirements How should it be done? 3 Regular functional requirement example (Webmail system): It should be possible to use HTML formatting in s Functional security requirement example: The system should check upon user registration that passwords are at least 8 characters long Non-functional security requirement example: All user input must be sanitized before being used in database queries How would you write a unit test for this?

4 Common security testing approaches Often difficult to craft e.g. unit tests from non-functional requirements Two common approaches: Test for known vulnerability types Attempt directed or random search of program state space to uncover the weird corner cases In today s lecture: Penetration testing (briefly) Fuzz testing or fuzzing Concolic testing 4

5 Penetration testing Manually try to break software Relies on human intuition and experience. Typically involves looking for known common problems. Can uncover problems that are impossible or difficult to find using automated methods but results completely dependent on skill of tester! 5

6 Fuzz testing Idea: Send semi-valid input to a program and observe its behavior. Black-box testing System Under Test (SUT) treated as a black-box The only feedback is the output and/or externally observable behavior of SUT. First proposed in a 1990 paper where completely random data was sent to 85 common Unix utilities in 6 different systems % crashed. Remember: Crash implies memory protection errors. Crashes are often signs of exploitable flaws in the program! 6

7 Fuzz testing architecture Fuzzing Framework Input Fuzzer Input Fuzzer generates inputs to SUT Dispatcher responsible for running SUT with input from fuzzer Assessor examines behavior of SUT to detect failures (i.e. signs of triggered bugs) Dispatcher Assessor SUT 7

8 Fuzzing components: Input generation Simplest method: Completely random Won t work well in practice Input deviates too much from expected format, rejected early in processing. Two common methods: Mutation based fuzzing Generation based fuzzing 8

9 Mutation based fuzzing Start with a valid seed input, and mutate it. Flip some bits, change value of some bytes. Programs that have highly structured input, e.g. XML, may require smarter mutations. Challenge: How to select appropriate seed input? If official test suites are available, these can be used. Generally mostly used for programs that take files as input. Trickier to do when interpretation of inputs depends on program state, e.g. network protocol parsers. (The way a message is handled depends on previous messages.) 9

10 Mutation based fuzzing Pros and Cons Easy to get started, no (or little) knowledge of specific input format needed. Typically yields low code coverage, inputs tend to deviate too much from expected format rejected by early sanity checks. int parse_input(char* data, size_t size) { int saved_checksum, computed_checksum; if(size < 4) return ERR_CODE; // First four bytes of data is CRC32 checksum saved_checksum = *((int*)data); Mutated inputs will always be rejected here! // Compute checksum for rest of data computed_checksum = CRC32(data + 4, size 4); // Error if checksums don t match if(computed_checksum!= saved_checksum) return ERR_CODE; 10 // Continue processing of data...

11 Mutation based fuzzing Pros and Cons Easy to get started, no (or little) knowledge of specific input format needed. Typically yields low code coverage, inputs tend to deviate too much from expected format rejected by early sanity checks. Hard to reach deeper parts of programs by random guessing int parse_record(char* data, int type) { switch(type) { case 0xFF0001: parse_type_a(data); break; case 0xFF0002: parse_type_b(data); break; Very unlikely to guess magic constants correctly. If seed only contains Type A records, parse_type_b will likely never be tested. case 0xFF0003: parse_type_c(data); break;... 11

12 Generation based fuzzing Idea: Use a specification of the input format (e.g. a grammar) to automatically generate semi-valid inputs Usually combined with various fuzzing heuristics that are known to trigger certain vulnerability types. Very long strings, empty strings Strings with format specifiers, extreme format strings %n%n%n%n%n%n%n%n%n%n%n%n%n%n%n %s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s % x Very large or small values, values close to max or min for data type 0x0, 0xffffffff, 0x7fffffff, 0x , 0xfffffffe Negative values where positive ones are expected 12

13 Generation based fuzzing Pros and Cons Input is much closer to the expected, much better coverage Can include models of protocol state machines to send messages in the sequence expected by SUT. Requires input format to be known. May take considerable time to write the input format grammar/specification. 13

14 Examples of generation based fuzzers (open source) SPIKE (2001): Early successful generation based fuzzer for network protocols. Designed to fuzz input formats consisting of blocks with fixed or variable size. E.g. [type][length][data] Peach (2006): Powerful fuzzing framework which allows specifying input formats in an XML format. Can represent complex input formats, but relatively steep learning curve. Sulley (2008): More modern fuzzing framework designed to be somewhat simpler to use than e.g. Peach. Also several commercial fuzzers, e.g. Codenomicon DEFENSICS, BeStorm, etc. 14

15 Fuzzing components: The Dispatcher Responsible for running the SUT on each input generated by fuzzer module. Must provide suitable environment for SUT. E.g. implement a client to communicate with a SUT using the fuzzed network protocol. SUT may modify environment (file system, etc.) Some fuzzing frameworks allow running SUT inside a virtual machine and restoring from known good snapshot after each SUT execution. 15

16 Fuzzing components: The Assessor Must automatically assess observed SUT behavior to determine if a fault was triggered. For C/C++ programs: Monitor for memory access violations, e.g. out-of-bounds reads or writes. Simplest method: Just check if SUT crashed. Problem: SUT may catch signals/exceptions to gracefully handle e.g. segmentation faults Difficult to tell if a fault, (which could have been exploitable with carefully crafted input), have occurred 16

17 Improving fault detection One solution is to attach a programmable debugger to SUT. Can catch signals/exceptions prior to being delivered to application. Can also help in manual diagnosis of detected faults by recording stack traces, values of registers, etc. However: All faults do not result in failures, i.e. a crash or other observable behavior (e.g. Heartbleed). An out-of-bounds read/write or use-after-free may e.g. not result in a memory access violation. Solution: Use a dynamic-analysis tool that can monitor what goes on under the hood Can potentially catch more bugs, but SUT runs (considerably) slower. Need more time for achieving the same level of coverage 17

18 AddressSanitizer Memory error checkers Two open source examples Applies instrumentation during compilation: Additional code is inserted in program to check for memory errors. Monitors all calls to malloc/new/free/delete can detect if memory is freed twice, used after free, out of bounds access of heap allocated memory, etc. Inserts checks that stack buffers are not accessed out of bounds Detects use of uninitialized variables etc Valgrind/Memcheck Applies instrumentation directly at the binary level during runtime does not need source code! Can detect similar problems as AddressSanitizer Applying instrumentation at the machine code level has some benefits works with any build environment, can instrument third-party libraries without source code, etc. But also comes at a cost; Runs slower than e.g. AddressSanitizer and can generally not detect out-of-bounds access to buffers on stack. Size of stack buffers not visible in machine code 18

19 Fuzzing web applications Errors are often on a higher abstraction level compared to e.g. simple coding errors leading to buffer overflows No hardware support for detecting faults i.e. program doesn t crash when something bad happens. Several additional challenges when fuzzing web applications: Determining inputs Assessing results (detecting faults) 19

20 Web application input vectors Inputs much more diverse than e.g. a file or a TCP connection: Form fields (name-value pairs and hidden fields) URL (names of pages and directories) HTTP headers Cookies Uploaded files AJAX How to identify inputs? Manually Using an automated crawler tries to visit all links and map out the web application 20

21 Web application input generation How to fuzz? XSS: Try including script tags SQL injection: Try supplying SQL-injection attack strings Try accessing resources that should require authentication 21

22 Assessing results of web app fuzzing Automatically detecting faults generally much harder than for native programs running on the CPU! Often requires abstract understanding of what is correct behavior For example: Detecting CSRF errors requires knowledge of what resources are supposed to be protected Difficult to cut the human out of the loop completely! 22

23 Assessing results of web app fuzzing Typical things to look for: Server errors Error messages (requires ad-hoc parsing of responses) Error codes (5xx errors, etc.) Signs of input (e.g. JavaScript) in responses Possible for reflected XSS Much harder for stored XSS! Error messages or SQL output indicating SQL-injection problems 23

24 Limitations of fuzz testing Many programs have an infinite input space and state space - Combinatorial explosion! Conceptually a simple idea, but many subtle practical challenges Difficult to create a truly generic fuzzing framework that can cater for all possible input formats. For best results often necessary to write a custom fuzzer for each particular SUT. (Semi)randomly generated inputs are very unlikely to trigger certain faults. 24

25 Limitations of fuzz testing Example from first lecture on vulnerabilities char buffer[100]; if(strlen(input) > 100) { printf( String too long! ); exit(1); } strcpy(buffer, input); The off-by-one error will only be detected if strlen(input) == 100 Very unlikely to trigger this bug using black-box fuzz testing! 25

26 Fuzzing outlook Mutation-based fuzzing can typically only find the low-hanging fruit shallow bugs that are easy to find Generation-based fuzzers almost invariably gives better coverage, but requires much more manual effort Current research in fuzzing attempts to combine the fire and forget nature of mutation-based fuzzing and the coverage of generation-based. Evolutionary fuzzing combines mutation with genetic algorithms to try to learn the input format automatically. Recent successful example is American Fuzzy Lop (AFL) Whitebox fuzzing generates test cases based on the control-flow structure of the SUT. Our next topic 26

27 Symbolic execution for program testing Recall symbolic execution: Encode a program path as a query to a SAT/SMT solver Have the solver find satisfying assignments In the lab you manually encoded paths as queries to a SMT solver. Of course also possible to perform this encoding automatically. 27

28 Concolic testing Idea: Combine concrete and symbolic execution Concolic execution (CONCrete and symbolic) Concolic execution workflow: 1. Execute the program for real on some input, and record path taken. 2. Encode path as query to SMT solver and negate one branch condition 3. Ask the solver to find new satisfying input that will give a different path Reported bugs are always accompanied by an input that triggers the bug (generated by SMT solver) Complete Reported bugs are always real bugs 28

29 Concolic testing small example void wheather(int temperature, unsigned int precipitation) { char forecast[5]; if(temperature > 0) { if(precipitation > 0) strcpy(forecast, rain ); else strcpy(forecast, nice ); } else { if(precipitation == 0) strcpy(forecast, cold ); else if(precipitation > 20) strcpy(forecast, blizzard ); else strcpy(forecast, snow ); } Buffer overflow!... 29

30 Concolic testing small example void wheather(int temperature, unsigned int precipitation ) { char forecast[5]; First round: if(temperature > 0) { if(precipitation > 0) strcpy(forecast, rain ); else strcpy(forecast, nice ); } else { if(precipitation == 0) strcpy(forecast, cold ); else if(precipitation > 20) strcpy(forecast, blizzard ); else strcpy(forecast, snow ); } Concrete inputs (arbitrary): temperature=1, precipitation=1 Recorded path constraints: temperature > 0 precipitation > 0 New path constraint: temperature > 0 (precipitation > 0) Solution from solver: temperature=1, precipitation=

31 Concolic testing small example void wheather(int temperature, unsigned int precipitation ) { char forecast[5]; Second round: if(temperature > 0) { if(precipitation > 0) strcpy(forecast, rain ); else strcpy(forecast, nice ); } else { if(precipitation == 0) strcpy(forecast, cold ); else if(precipitation > 20) strcpy(forecast, blizzard ); else strcpy(forecast, snow ); } Note: precipitation is unconstrained no need to care about later branch conditions when... negating an earlier condition. Concrete inputs: temperature=1, precipitation=0 Recorded path constraints: temperature > 0 (precipitation > 0) New path constraint: (temperature > 0) Solution from solver: temperature=0, precipitation=0 31

32 Concolic testing small example void wheather(int temperature, unsigned int precipitation ) { char forecast[5]; Third round: if(temperature > 0) { if(precipitation > 0) strcpy(forecast, rain ); else strcpy(forecast, nice ); } else { if(precipitation == 0) strcpy(forecast, cold ); else if(precipitation > 20) strcpy(forecast, blizzard ); else strcpy(forecast, snow ); } Concrete inputs: temperature=0, precipitation=0 Recorded path constraints: (temperature > 0) precipitation = 0 New path constraint: (temperature > 0) (precipitation = 0) Solution from solver: temperature=0, precipitation=

33 Concolic testing small example void wheather(int temperature, unsigned int precipitation ) { char forecast[5]; Fourth round: if(temperature > 0) { if(precipitation > 0) strcpy(forecast, rain ); else strcpy(forecast, nice ); } else { if(precipitation == 0) strcpy(forecast, cold ); else if(precipitation > 20) strcpy(forecast, blizzard ); else strcpy(forecast, snow ); }... Concrete inputs: temperature=0, precipitation=1 Recorded path constraints: (temperature > 0) (precipitation = 0) (precipitation > 20) New path constraint: (temperature > 0) (precipitation = 0) precipitation > 20 Solution from solver: temperature=0, precipitation=21 33

34 Concolic testing small example void wheather(int temperature, unsigned int precipitation ) { char forecast[5]; Fifth round: if(temperature > 0) { if(precipitation > 0) strcpy(forecast, rain ); else strcpy(forecast, nice ); } else { if(precipitation == 0) strcpy(forecast, cold ); else if(precipitation > 20) strcpy(forecast, blizzard ); else strcpy(forecast, snow ); } Concrete inputs: temperature=0, precipitation=21 Recorded path constraints: (temperature > 0) (precipitation = 0) precipitation > 20 Bug found!... 34

35 Challenges with concolic testing: Path explosion Number of paths increase exponentially with number of branches Most real-world programs have an infinite state space! For example, number of loop iterations may depend on size of input Not possible to explore all paths Need a strategy to explore interesting parts of the program 37

36 Challenges with concolic testing: Path explosion Depth-first search (as in the first example) will easily get stuck in one part of the program May e.g. keep exploring the same loop with more and more iterations Breadth-first search will take a very long time to reach deep states May take forever to reach the buggy code Try smarter ways of exploring the program state space May want to try to run loops many times to uncover possible buffer overflows but also want to maximize coverage of different parts of the program 38

37 Generational search ( whitebox fuzzing ) The Microsoft SAGE system implements whitebox fuzzing Performs concolic testing, but prioritizes paths based on how much they improve coverage Results can be assessed similar to black-box fuzzing (with dynamic analysis tools, etc.) Search algorithm outline: 1. Run program on concrete seed input and record path constraint 2. For each branch condition: I. Negate condition and keep the earlier conditions unchanged. II. III. Have SMT solver generate new satisfying input and run program with that input. Assign input a score based on how much it improves coverage (I.e. how much previously unseen code will be executed with that input) IV. Store input in a global worklist sorted on score 3. Pick input at head of worklist and repeat. 39

38 Rationale for whitebox fuzzing Coverage based heuristic avoids getting stuck as in DFS If one test case executes the exact same code as in a previous test case its score will be 0 Moved to end of worklist Probably never used again Gives sparse but more diverse search of the paths of the program Implements a form of greedy search heuristic For example, loops may only be executed once! Only generates input close to the seed input (cf. mutation-based fuzzing) Will try to explore the weird corner cases in code exercised by seed input Choice of seed input important for good coverage 45

39 Rationale for whitebox fuzzing Has proven to work well in practice Used in production at Microsoft to test e.g. Windows, Office, etc. prior to release Has uncovered many serious vulnerabilities that was missed by other approaches (black-box fuzzing, static analysis, etc.) Interestingly, SAGE works directly at the machine-code level Note: Source code not needed for concolic execution sufficient to collect constraints from one concrete sequence of machine-code instructions. Avoids hassle with different build environments, third-party libraries, programs written in different languages etc. but sacrifices some coverage due to additional approximations needed when working on machine code 46

40 Limitations of concolic testing The success of concolic testing is due to the massive improvement in SAT/SMT solvers during the last two decades. Main bottleneck is still often the solvers. Black-box fuzzing can perform a much larger number of test cases per time unit may be more time efficient for shallow bugs. Solving SAT/SMT problems is NP-complete. Solvers like e.g. Z3 use various tricks to speed up common cases but may take unacceptably long time to solve certain path constraints. Solving SMT queries with non-linear arithmetic (multiplication, division, modulo, etc.) is an undecidable problem in general. But since programs typically use fixed-precision data types, non-linear arithmetic is decidable in this special case. 47

41 Limitations of concolic testing If program uses any kind of cryptography, symbolic execution will typically fail. Consider previous checksum example: CRC32 is linear and reversible solver can repair checksum if rest of data is modified.... // Compute checksum for rest of data computed_checksum = CRC32(data + 4, size 4); // Error if checksums don t match if(computed_checksum!= saved_checksum) return ERR_CODE;... What if program used e.g. md5 or SHA256 here instead? Solver would get stuck trying to solve this constraint! Generation-based fuzzing could handle this without problem! 48

42 Greybox fuzzing Probability of hitting a deep level of the code decreases exponentially with the depth of the code for mutation based fuzzing. Similarly, the time required for solving an SMT query is high, and increases exponentially with the depth of the path constraint. Black-box fuzzing is too dumb and whitebox fuzzing may be too smart Idea of greybox fuzzing is to find a sweet spot in between. if(condtion1) if(condtion2) if(condtion3) if(condtion4) bug(); Mutational fuzzer would need to guess correct values of all four condtions in one go to reach bug! 49

43 Greybox fuzzing Instead of recording full path constraint (as in whitebox fuzzing), record light-weight coverage information to guide fuzzing. Use evolutionary algorithms to learn input format. American Fuzzy Lop (AFL) is considered the current state-of-the art in fuzzing. Performs regular mutation-based fuzzing (using several different strategies) and measures code coverage. Every generated input that resulted in any new coverage is saved and later re-fuzzed This extremely simple evolutionary algorithm allows AFL to gradually learn how to reach deeper parts of the program. Also highly optimized for speed can reach several thousand test cases per second. This often beats smarter (and slower) methods like whitebox fuzzing! Has found hundreds of serious vulnerabilities in open-source programs! 50

44 Conclusions Fuzzing Black-box testing method. Semi-random input generation. Despite being a brute-force, somewhat ad-hoc approach to security testing, experience has shown that it improves security in practice. Concolic testing White-box testing method. Input generated from control-structure of code to systematically explore different paths of the program. Some in-house and academic implementations exist, but some challenges left to solve before widespread adoption. Greybox fuzzing Coverage-guided semi-random input generation. High speed sometimes beats e.g. concolic testing, but shares some limitations with mutation-based fuzzing (e.g. magic constants, checksums) 51

45 Conclusions Test cases generated automatically, either semi-randomly or from structure of code Test cases not based on requirements Pro: Not biased by developers view of how things should work, can uncover unsound assumptions or corner cases not covered by specification. Con: Fuzzing and concolic testing mostly suited for finding implementation errors, e.g. buffer overflows, arithmetic overflows, etc. Generally hard to test for high-level errors in requirements and design using these methods 52

46 Conclusions Different methods are good at finding different kinds of bugs, but none is a silver bullet. Fuzzing cannot be the only security assurance method used! Static analysis Manual reviews (code, design documents, etc.) Regular unit/integration/system testing! 53